Builds the conformance input-log fixture inside its cases

[johnnyt]

What

StatifierPersistence.Testing.StorageConformance registered a setup that
inserted the input-log fixture run. ExUnit runs setup callbacks in
definition order and the template's are defined where a host writes use,
so a host whose adapter needs a per-test binding established before any
write - a session parameter, a connection-scoped setting, a sandbox
checkout - found a row already written by the time its own callback ran,
with nothing in the docs saying why or what to do instead.

Both halves of sp-flj:

1. The fixture is lazy. input_log_run/2 is called inside each of the
   three cases that need it, rather than from a setup. The only callback
   the template registers now opens a handle and, where the adapter exports
   isolate/1, isolates it - it writes nothing.
2. The moduledoc states the contract. A setup a host needs before any
   write goes above the use; a setup written below it is guaranteed
   only that no row has been written yet, not that nothing has run. The
   worked example is in the moduledoc.

The test

StatifierPersistence.Storage.SetupOrderingConformanceTest is the
host-shaped case: a setup defined deliberately BELOW the use, so it
runs after every callback the template registers, asserting the handle is
open (the template's setup ran first) and the input-log fixture run is
absent. Registered for the whole module, so the assertion holds over every
case the template generates for the input-log adapter, not over one.

Both new declarations carry a # sabotage: note, and both mutations were
run:

• restoring the eager setup in the template -> 38 of 38 cases in the
  new module red, the fixture run already inserted when the host-shaped
  callback ran;
• dropping Storage.insert_run/4 from input_log_run/2 -> 5 cases
  red, reporting :run_not_found where a run was expected.

Both reverted from a copy taken before the edit; the gate is green on the
reverted tree.

Verification

• Full mix quality green on this HEAD: format, compile (warnings as
  errors), credo, dialyzer, deps, 520 of 520 tests, 95.7% coverage.
  mix quality.verify attests the run (scope all, no profile).
• The input-log cases pass on both adapters: InputLogConformanceTest
  (in-memory) and EctoConformanceTest (real Postgres) run green, and a
  --trace run shows the input-log case names still generated under the
  Ecto adapter.
• Only one setup remains in the template (the use-registered one at
  storage_conformance.ex), and its body is Storage.new/2 plus the
  optional isolate/1.

Provenance

No file outside the bead's map was touched. changelog.d/sp-flj.md under
Fixed: the template ships in lib/ and is public surface an adapter
package binds against, so the ordering change is visible to someone who
only uses the public API.

Refs: sp-flj